Deep Learning Model Accurately Detects Ankylosing Spondylitis on MRI Save

Artificial intelligence (AI) continues to move from research labs into rheumatology clinics, promising to reshape how inflammatory diseases are diagnosed and monitored. Ankylosing spondylitis (AS), in particular, poses a diagnostic challenge—its early imaging findings are subtle, and radiographic changes may take years to appear and require specific musculoskeletal radiology expertise which may not be universally available. A new deep learning model, ASembleNet, offers a glimpse into how AI could bridge this diagnostic gap.

A very interesting abstract (#2683) at #ACR25 highlights one such model developed by a multidisciplinary team using multi-institutional, publicly available MRI datasets. ASembleNet integrates convolutional neural networks with transformer-based architectures to classify AS on MRI scans. The model achieved impressive results: 99% accuracy on coronal and 98% on axial MRI, outperforming established deep learning frameworks such as Xception and ResNet50. External validation with expert-reviewed cases from the Assessment of Spondyloarthritis International Society (ASAS) confirmed its generalizability, and Grad-CAM visualization demonstrated that the model focused on clinically relevant regions of interest—an essential step toward interpretability.

For rheumatologists, tools like ASembleNet may soon offer decision-support in ambiguous or early-stage cases, where radiologists and clinicians struggle to distinguish inflammatory from mechanical back pain. Such algorithms could streamline image triage, reduce diagnostic delay, and support early treatment decisions that alter long-term outcomes.

As more AI tools enter rheumatology, clinicians must evolve from passive users to informed partners in their development and application. Understanding whether an algorithm was trained on public or proprietary datasets, and whether its architecture is open for clinical use or independent validation, is essential. These details shape not only the credibility of AI findings but also their equitable adoption and integration across diverse healthcare settings.

As rheumatologists, we need to start actively adapting to the era of AI-assisted imaging—working alongside data scientists, contributing to open datasets, and advocating for transparency in model design and dissemination. Only through such collaboration can we ensure that emerging technologies augment, rather than replace, clinical judgment.

ASembleNet marks a step forward, but its greatest value will depend on how openly and responsibly such models are shared and disseminated. The abstract left questions unanswered about accessibility—a critical factors if we want AI to be adapted vastly.